fairtalking-second-work / scripts /analysis /draw_cta_framework_v3.py
huahua123313's picture
Add files using upload-large-folder tool
eea47ad verified
Raw
History Blame Contribute Delete
16.7 kB
"""Render the CTA (A1_full) framework — v3, complete training-time view.
Improvements over v2:
* Cross-attention internals shown (Q from tgt-modality, K/V from src-modality).
* All FIVE losses drawn as separate dashed back-arrows, color-coded.
* detach() barrier explicit (red ⊥ symbol on the s_asym→classifier path).
* is_real mask shown on loss_av/loss_va (★ REAL only badges).
* Real-sample numbers from trace_cta_training.log are annotated next to
L_AV / L_VA / s_asym / score boxes.
Output: outputs/analysis/figs_framework/cta_framework_v3.{png,pdf}
Run from project root:
python3 scripts/analysis/draw_cta_framework_v3.py
"""
from __future__ import annotations
from pathlib import Path
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from matplotlib.patches import FancyArrowPatch, FancyBboxPatch, Rectangle
OUT_DIR = Path("outputs/analysis/figs_framework")
OUT_DIR.mkdir(parents=True, exist_ok=True)
# ============================================================
# Palette
# ============================================================
C_REAL = "#1E8449"
C_FAKE = "#C0392B"
C_SHARED = "#5D6D7E"
C_AUDIO = "#B7950B"
C_DATA_REAL_FILL = "#E8F8F0"
C_DATA_FAKE_FILL = "#FBEAEA"
C_DATA_AUDIO_FILL = "#FCF3CF"
C_BB_FILL = "#EAF2F8"
C_BB_EDGE = "#5499C7"
C_CORE_FILL = "#FDEBD0"
C_CORE_EDGE = "#D35400"
C_CORE_TEXT = "#7B241C"
C_OUT_FILL = "#E8DAEF"
C_OUT_EDGE = "#7D3C98"
C_PANEL_LOSS = "#FEF9E7"
C_PANEL_LOSS_EDGE = "#B7950B"
# Loss arrow colors — one per loss
C_L_AV = "#1E8449" # green
C_L_VA = "#138D75" # teal
C_L_ASYM = "#7D3C98" # purple
C_L_CLS = "#1F618D" # blue
C_L_AUX = "#C0392B" # red
# ============================================================
# Helpers
# ============================================================
def box(ax, cx, cy, w, h, text, fill, edge, *,
fontsize=10, weight="normal", lw=1.5, padding=0.4, star=False,
text_color="black"):
rect = FancyBboxPatch(
(cx - w/2, cy - h/2), w, h,
boxstyle=f"round,pad={padding}",
facecolor=fill, edgecolor=edge, linewidth=lw,
)
ax.add_patch(rect)
label = ("★ " + text) if star else text
ax.text(cx, cy, label, ha="center", va="center",
fontsize=fontsize, weight=weight, color=text_color)
def arrow(ax, x1, y1, x2, y2, *, color=C_SHARED, lw=1.6, dashed=False,
mutation=14, alpha=1.0, connectionstyle="arc3,rad=0"):
ls = (0, (5, 3)) if dashed else "-"
a = FancyArrowPatch(
(x1, y1), (x2, y2),
arrowstyle="->", mutation_scale=mutation,
color=color, linewidth=lw, linestyle=ls, alpha=alpha,
shrinkA=2, shrinkB=2, connectionstyle=connectionstyle,
)
ax.add_patch(a)
def small(ax, cx, cy, text, *, fontsize=8.5, color="#566573", italic=True,
ha="center", weight="normal"):
style = "italic" if italic else "normal"
ax.text(cx, cy, text, ha=ha, va="center",
fontsize=fontsize, style=style, color=color, weight=weight)
def detach_barrier(ax, x, y, length=2.0, color="#C0392B"):
"""Draw a small ⊥ to indicate detach() / gradient barrier."""
ax.plot([x - length/2, x + length/2], [y, y],
color=color, linewidth=2.5, zorder=5)
ax.plot([x, x], [y - length/3, y + length/3],
color=color, linewidth=2.5, zorder=5)
# ============================================================
# Canvas — wider to fit cross-attention internals
# ============================================================
fig, ax = plt.subplots(figsize=(20, 13))
ax.set_xlim(0, 100)
ax.set_ylim(0, 100)
ax.set_aspect("equal")
ax.axis("off")
# ---- Title ----
ax.text(50, 96.5, "CTA Framework — Forward Pass + Five Losses + Gradient Routing",
ha="center", va="center", fontsize=17, weight="bold")
small(ax, 50, 93.5,
"Solid arrows = forward. Dashed arrows = gradient (color-coded by loss). "
"★ marks our four innovations. ⊥ marks detach() barrier.",
fontsize=9.5, italic=True, color="#566573")
# ---- tiny legend (forward / shared / detach) ----
ax.plot([4, 7], [89.5, 89.5], color=C_REAL, lw=2.4)
ax.text(7.4, 89.5, "real path", fontsize=8.5, va="center", color=C_REAL, weight="bold")
ax.plot([15, 18], [89.5, 89.5], color=C_FAKE, lw=2.4, linestyle=(0, (5, 3)))
ax.text(18.4, 89.5, "fake path", fontsize=8.5, va="center", color=C_FAKE, weight="bold")
ax.plot([26, 29], [89.5, 89.5], color=C_SHARED, lw=2.4)
ax.text(29.4, 89.5, "shared", fontsize=8.5, va="center", color=C_SHARED)
# ============================================================
# INPUTS (y ≈ 84)
# ============================================================
y_in = 85
box(ax, 10, y_in, 14, 4.5,
"Real video",
C_DATA_REAL_FILL, C_REAL, fontsize=10)
box(ax, 26, y_in, 14, 4.5,
"Fake video",
C_DATA_FAKE_FILL, C_FAKE, fontsize=10)
box(ax, 49, y_in, 22, 4.5,
"Audio ★ ALWAYS REAL",
C_DATA_AUDIO_FILL, C_AUDIO, fontsize=10, weight="bold",
text_color="#7E5109")
small(ax, 10, y_in - 3.2, "label = 0", fontsize=8, italic=False, color=C_REAL)
small(ax, 26, y_in - 3.2, "label = 1", fontsize=8, italic=False, color=C_FAKE)
small(ax, 49, y_in - 3.2, "(generator only synthesizes video)", fontsize=8, italic=True)
# ============================================================
# BACKBONES (y ≈ 75)
# ============================================================
y_bb = 75
box(ax, 18, y_bb, 26, 5,
"VideoMAE-base (freeze 70 %)\nv.tokens (B,1568,768) , v.pooled (B,768)",
C_BB_FILL, C_BB_EDGE, fontsize=9.5)
box(ax, 49, y_bb, 22, 5,
"Wav2Vec2-base (freeze 80 %)\na.tokens (B,127,768) , a.pooled (B,768)",
C_BB_FILL, C_BB_EDGE, fontsize=9.5)
# input → backbone arrows
arrow(ax, 10, y_in - 4.5, 14, y_bb + 2.5, color=C_REAL, lw=2.0)
arrow(ax, 26, y_in - 4.5, 22, y_bb + 2.5, color=C_FAKE, lw=2.0, dashed=True)
arrow(ax, 49, y_in - 4.5, 49, y_bb + 2.5, color=C_SHARED, lw=2.0)
# ============================================================
# CROSS-MODAL PREDICTORS (y ≈ 60) with cross-attention insets
# ============================================================
y_pr = 60
# left predictor (A→V)
predA = FancyBboxPatch(
(5, y_pr - 5), 30, 10,
boxstyle="round,pad=0.4",
facecolor=C_CORE_FILL, edgecolor=C_CORE_EDGE, linewidth=2.5,
)
ax.add_patch(predA)
ax.text(20, y_pr + 3, r"★ $f_{A \to V}$ predictor",
ha="center", fontsize=11, weight="bold", color=C_CORE_TEXT)
ax.text(20, y_pr + 1.4, "Transformer Decoder × 4",
ha="center", fontsize=9, color=C_CORE_TEXT)
# attention sub-block
ax.text(13, y_pr - 1.2, "Q", ha="center", fontsize=9, color="black", weight="bold")
ax.text(20, y_pr - 1.2, "K, V", ha="center", fontsize=9, color="black", weight="bold")
ax.text(13, y_pr - 2.4, "(v.tokens)", ha="center", fontsize=7.5, style="italic", color="#566573")
ax.text(20, y_pr - 2.4, "(a.tokens)", ha="center", fontsize=7.5, style="italic", color="#566573")
ax.text(27, y_pr - 1.8, "→ v_pred", ha="center", fontsize=9.5, color=C_CORE_TEXT, weight="bold")
small(ax, 20, y_pr - 4.2, "trained on REAL pairs only",
fontsize=8.2, italic=True, color=C_REAL, weight="bold")
# right predictor (V→A)
predB = FancyBboxPatch(
(40, y_pr - 5), 30, 10,
boxstyle="round,pad=0.4",
facecolor=C_CORE_FILL, edgecolor=C_CORE_EDGE, linewidth=2.5,
)
ax.add_patch(predB)
ax.text(55, y_pr + 3, r"★ $f_{V \to A}$ predictor",
ha="center", fontsize=11, weight="bold", color=C_CORE_TEXT)
ax.text(55, y_pr + 1.4, "Transformer Decoder × 4",
ha="center", fontsize=9, color=C_CORE_TEXT)
ax.text(48, y_pr - 1.2, "Q", ha="center", fontsize=9, color="black", weight="bold")
ax.text(55, y_pr - 1.2, "K, V", ha="center", fontsize=9, color="black", weight="bold")
ax.text(48, y_pr - 2.4, "(a.tokens)", ha="center", fontsize=7.5, style="italic", color="#566573")
ax.text(55, y_pr - 2.4, "(v.tokens)", ha="center", fontsize=7.5, style="italic", color="#566573")
ax.text(62, y_pr - 1.8, "→ a_pred", ha="center", fontsize=9.5, color=C_CORE_TEXT, weight="bold")
small(ax, 55, y_pr - 4.2, "trained on REAL pairs only",
fontsize=8.2, italic=True, color=C_REAL, weight="bold")
# arrows backbone → predictor (only the dominant tgt arrows; src arrows shown as ghosts)
arrow(ax, 18, y_bb - 2.5, 13, y_pr + 5, color=C_SHARED, lw=1.6)
arrow(ax, 49, y_bb - 2.5, 20, y_pr + 5, color=C_SHARED, lw=1.0, alpha=0.5)
arrow(ax, 49, y_bb - 2.5, 48, y_pr + 5, color=C_SHARED, lw=1.6)
arrow(ax, 18, y_bb - 2.5, 55, y_pr + 5, color=C_SHARED, lw=1.0, alpha=0.5)
# ============================================================
# RESIDUALS (y ≈ 47) — per-sample MSE
# ============================================================
y_res = 47
box(ax, 13, y_res, 22, 5,
"$L_{A \\to V}$ = MSE(v_pred, v.tokens)\n[per-sample scalar]",
C_CORE_FILL, C_CORE_EDGE, fontsize=10, lw=2.0)
box(ax, 55, y_res, 22, 5,
"$L_{V \\to A}$ = MSE(a_pred, a.tokens)\n[per-sample scalar]",
C_CORE_FILL, C_CORE_EDGE, fontsize=10, lw=2.0)
arrow(ax, 13, y_pr - 5, 13, y_res + 2.5, color=C_SHARED, lw=1.4)
arrow(ax, 62, y_pr - 5, 55, y_res + 2.5, color=C_SHARED, lw=1.4)
# real / fake values from trace_cta_training.log
small(ax, 13, y_res - 4.8,
"real: 0.001 fake: 0.012 (×12)",
fontsize=8, italic=False, color="#1B4F72")
small(ax, 55, y_res - 4.8,
"real: 0.0004 fake: 0.0023 (×6)",
fontsize=8, italic=False, color="#1B4F72")
# ============================================================
# ASYMMETRY SCORE (y ≈ 36)
# ============================================================
y_asym = 36
box(ax, 34, y_asym, 50, 5,
r"$s_{\rm asym}$ = $L_{V \to A}$ − $L_{A \to V}$ "
r"$L_{\rm total}$ = $L_{V \to A}$ + $L_{A \to V}$",
C_CORE_FILL, C_CORE_EDGE, fontsize=10.5, lw=2.5, star=True,
text_color=C_CORE_TEXT, weight="bold")
small(ax, 34, y_asym - 3.6,
"real $s_{\\rm asym}$ ≈ -0.0007 fake $s_{\\rm asym}$ ≈ -0.010 (gap × 14)",
fontsize=8, italic=False, color="#1B4F72")
arrow(ax, 13, y_res - 2.5, 25, y_asym + 2.5, color=C_SHARED, lw=1.4)
arrow(ax, 55, y_res - 2.5, 43, y_asym + 2.5, color=C_SHARED, lw=1.4)
# ============================================================
# DETACH BARRIER + CLASSIFIER HEAD
# ============================================================
y_cls = 25.5
# vertical arrow from s_asym down to classifier, with detach() barrier
ax.annotate("", xy=(34, y_cls + 3.5), xytext=(34, y_asym - 2.5),
arrowprops=dict(arrowstyle="->", lw=1.6, color=C_SHARED))
detach_barrier(ax, 34, y_asym - 4.6, length=3.0, color="#C0392B")
ax.text(36.5, y_asym - 4.6, ".detach()",
fontsize=9, color="#C0392B", weight="bold", style="italic", va="center")
box(ax, 34, y_cls, 56, 6,
"Classifier Head MLP (1538 → 256 → 1)\n"
r"input = concat[ v.pooled , a.pooled , $s_{\rm asym}$ , $L_{\rm total}$ ]",
C_BB_FILL, C_BB_EDGE, fontsize=10)
# pooled features feed in from the side (subtle long curves)
arrow(ax, 18, y_bb - 2.5, 8, y_cls + 1, color=C_SHARED, lw=0.7, alpha=0.45,
connectionstyle="arc3,rad=-0.3")
arrow(ax, 49, y_bb - 2.5, 60, y_cls + 1, color=C_SHARED, lw=0.7, alpha=0.45,
connectionstyle="arc3,rad=0.3")
small(ax, 11, y_cls + 4, "v.pooled", fontsize=7.5, color="#7F8C8D")
small(ax, 57, y_cls + 4, "a.pooled", fontsize=7.5, color="#7F8C8D")
# ============================================================
# OUTPUT
# ============================================================
y_out = 14.5
box(ax, 34, y_out, 28, 5,
"score = sigmoid(logit) ∈ [0, 1]\nhigh ⇒ fake real: 0.0006 fake: 0.997",
C_OUT_FILL, C_OUT_EDGE, fontsize=10, weight="bold")
arrow(ax, 34, y_cls - 3, 34, y_out + 2.5, color=C_SHARED, lw=1.6)
# ============================================================
# FIVE LOSS BACK-ARROWS (color-coded; dashed; with badges)
# ============================================================
def loss_arrow(x1, y1, x2, y2, color, lw=1.6, rad=0.0):
a = FancyArrowPatch(
(x1, y1), (x2, y2),
arrowstyle="->", mutation_scale=14,
color=color, linewidth=lw, linestyle=(0, (4, 3)),
shrinkA=3, shrinkB=3,
connectionstyle=f"arc3,rad={rad}",
)
ax.add_patch(a)
# loss_av: green dashed back-arrow, L_AV → A→V predictor (only this loss)
loss_arrow(13, y_res + 2.5, 13, y_pr - 5, C_L_AV, lw=2.0, rad=-0.4)
ax.text(7.5, 51.5, r"$loss_{av}$",
fontsize=10, color=C_L_AV, weight="bold")
ax.text(7.5, 49.6, "★ REAL only",
fontsize=7.5, color=C_L_AV, weight="bold", style="italic")
ax.text(7.5, 47.7, "(is_real mask)",
fontsize=7, color="#566573", style="italic")
# loss_va: teal dashed back-arrow, L_VA → V→A predictor
loss_arrow(55, y_res + 2.5, 55, y_pr - 5, C_L_VA, lw=2.0, rad=0.4)
ax.text(66, 51.5, r"$loss_{va}$",
fontsize=10, color=C_L_VA, weight="bold")
ax.text(66, 49.6, "★ REAL only",
fontsize=7.5, color=C_L_VA, weight="bold", style="italic")
ax.text(66, 47.7, "(is_real mask)",
fontsize=7, color="#566573", style="italic")
# loss_asym: purple, from s_asym box back to BOTH predictors
loss_arrow(20, y_asym + 2.5, 20, y_pr - 5, C_L_ASYM, lw=1.4, rad=-0.45)
loss_arrow(48, y_asym + 2.5, 48, y_pr - 5, C_L_ASYM, lw=1.4, rad=0.45)
ax.text(34, y_asym + 6.5,
r"$loss_{asym}$ = ReLU($s_f - s_r$) → margin (BOTH)",
ha="center", fontsize=9.5, color=C_L_ASYM, weight="bold")
# loss_cls: blue, from output back to classifier; STOPS at detach() barrier
# (visualized by the arrow ending at y_cls and a separate ⊥ to predictors)
loss_arrow(34, y_out + 2.5, 34, y_cls - 3, C_L_CLS, lw=2.0, rad=0.0)
ax.text(46, y_cls + 4.5,
r"$loss_{cls}$ = BCE(score, label) → classifier (BOTH)",
ha="center", fontsize=9.5, color=C_L_CLS, weight="bold")
# explicit "blocked by detach" annotation on the path classifier→predictor
ax.text(40, y_asym - 4.6, " ← BCE gradient blocked here",
fontsize=7.5, color="#C0392B", style="italic", va="center")
# loss_aux: red, only on fake (drawn as a small badge on the side)
ax.text(82, 38, r"$loss_{aux}$",
fontsize=10, color=C_L_AUX, weight="bold")
ax.text(82, 36.5,
r"= MSE($s_{\rm asym}$, $s_{\rm asym, alt\_gen}$)",
fontsize=8, color=C_L_AUX, style="italic")
ax.text(82, 35,
"★ paired FAKEs (cross-generator)",
fontsize=7.5, color=C_L_AUX, weight="bold")
loss_arrow(82, 39.2, 70, y_pr - 4, C_L_AUX, lw=1.0, rad=0.4)
# ============================================================
# Right-side LEGEND PANEL of the 5 losses
# ============================================================
lp_w, lp_h = 28, 24
lp_cx, lp_cy = 86, 14
lp = FancyBboxPatch(
(lp_cx - lp_w/2, lp_cy - lp_h/2), lp_w, lp_h,
boxstyle="round,pad=0.5",
facecolor=C_PANEL_LOSS, edgecolor=C_PANEL_LOSS_EDGE, linewidth=2,
)
ax.add_patch(lp)
ax.text(lp_cx, lp_cy + lp_h/2 - 1.6, "Five Losses & gradient routing",
ha="center", fontsize=11, weight="bold")
rows = [
(r"$loss_{av}$", C_L_AV, "REAL only", "→ A→V predictor + backbones"),
(r"$loss_{va}$", C_L_VA, "REAL only", "→ V→A predictor + backbones"),
(r"$loss_{asym}$", C_L_ASYM, "BOTH (margin)", "→ both predictors (weak)"),
(r"$loss_{cls}$", C_L_CLS, "BOTH (BCE)", "→ classifier head only"),
(r"$loss_{aux}$", C_L_AUX, "paired FAKEs", "→ both predictors (cross-gen)"),
]
ay = lp_cy + lp_h/2 - 4
for name, color, mask, dest in rows:
ax.plot([lp_cx - lp_w/2 + 1.0, lp_cx - lp_w/2 + 3.5],
[ay + 0.3, ay + 0.3],
color=color, lw=2.2, linestyle=(0, (4, 3)))
ax.text(lp_cx - lp_w/2 + 4.0, ay + 0.3, name,
fontsize=9, color=color, weight="bold")
ax.text(lp_cx - lp_w/2 + 9, ay + 0.3, mask,
fontsize=7.5, color="#566573", style="italic")
ax.text(lp_cx - lp_w/2 + 1.0, ay - 1.5, dest,
fontsize=7.2, color="#34495E")
ay -= 3.8
ax.text(lp_cx, lp_cy - lp_h/2 + 1.5,
r"Total = 1·$loss_{av}$ + 1·$loss_{va}$ + 0.5·$loss_{asym}$"
"\n "
r"+ 1·$loss_{cls}$ + 0.1·$loss_{aux}$",
ha="center", fontsize=8.5, weight="bold")
# ============================================================
# Save
# ============================================================
out_png = OUT_DIR / "cta_framework_v3.png"
out_pdf = OUT_DIR / "cta_framework_v3.pdf"
plt.tight_layout(pad=0.5)
plt.savefig(out_png, dpi=240, bbox_inches="tight", facecolor="white")
plt.savefig(out_pdf, bbox_inches="tight", facecolor="white")
plt.close(fig)
print(f"[draw] wrote {out_png}")
print(f"[draw] wrote {out_pdf}")